Prediction of Esports Game Results Using Early Game Datasets

Abstract:
Large e-sports games are becoming prevailing at an unprecedented around the world in recent years. Thus, an increasing amount of effort is put into game-predictions for worldwide audiences. However, most of the game prediction results are restricted by both the high uncertainty at the early-game stage, and the hardship of interpreting them for audiences. In this paper, we mainly focus on predicting the game results for one of the most popular games, League of Legends, using the first ten minute in-game features. This research uses Active Learning as the main method to evaluate the dataset that contains approximately ten thousand high-Elo ranked games. Then, the research compares the results of prediction based on our active learning algorithms and the random selection methods. The research experiment finally presents that the prediction result of the active learning model is similar to that of the traditional machine learning methods and outweigh the accuracy of random selection. The paper can conclude that we may save a lot of human-needed work and use a relatively small amount of initial datasets to generate more accurate game predictions by implementing active learning in large data-based e-sports matches.